Research on the Application of Deep Reinforcement Learning in Road Network Control

In response to the escalating issue of expressway network congestion in urban agglomerations and the inadequacy of traditional control methods in adapting to-dynamic and complex road networks, this study aims to explore the application potential of multi-agent deep reinforcement learning in coordinated regional road network control. The research proposes a traffic control model based on the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm, which abstracts multiple key nodes in the road network into independent agents. These agents employ a centralized training and decentralized execution approach, where they learn collaborative strategies using shared global information during training, but make independent decisions based on local observations during execution. Through interaction with the SUMO (Simulation of Urban Mobility) traffic simulation platform, the agents learn to dynamically adjust lane opening/closing statuses to optimize overall road network traffic efficiency. Simulation results demonstrate that, compared to scenarios with no control or traditional single-lane control, the proposed MADDPG model significantly improves average road network speed and effectively alleviates traffic congestion, thereby validating the feasibility and superiority of this method in achieving intelligent and coordinated road network traffic management. This research provides new insights and empirical support for the application of deep reinforcement learning techniques in addressing complex expressway network problems in urban agglomerations.